Skeleton-Based Action Recognition

Skeleton-based action recognition studies how to infer activities from the positions and movements of body joints over time. Common approaches use graph convolutional networks to model relationships between joints, alongside temporal transformers, state-space models, and geometry-aware methods to capture motion patterns at different scales. Self-supervised learning, generative augmentation, and alignment with language or video features aim to improve recognition when labels are scarce or actions are unseen; methods are often evaluated on benchmarks such as NTU RGB+D and PKU-MMD. A central challenge is reliable transfer beyond controlled benchmarks, where camera changes, missing joints, and different skeleton formats can undermine performance, while richer visual or semantic cues may improve recognition at a cost to privacy or skeleton-only deployment.